On the computation of linearly constrained stationary points (Q1331099)

Various gradient-type algorithms are compared for the numerical minimization of a function defined on a Hilbert space, subject to linear constraints. One wants to find an element \(x\) of a Hilbert space \(H\), such that a given function \(f: H\to \mathbb{R}\) is minimized subject to \(Bx= a\), where \(a\in V\) and \(B: H\to V\) is a bounded linear operator; \(V\) is also a Hilbert space. The convergence properties of three algorithms, viz. 1. the Arrow-Hurwitz gradient method, 2. the projected gradient method, and 3. the perturbed projected gradient method are studied. In the paper the latter method is developed and emphasized. Theorems on convergence rates are stated and proved. An assumption in these theorems is that the operator \(BB^*\) is strictly positive. Later on results are presented when this condition does not hold. For the case that \(f\) is a quadratic function of \(x\), special results are obtained. The paper concludes with two numerical examples. In the first one \(H\) is finite-dimensional (in fact \(H= \mathbb{R}^ 4\)) and in the second one there is a continuous problem (the problem description includes a simple differential equation) which has been discretized to apply the algorithm. The results are compared to the analytically known solution.